Emotion Forecasting: A Transformer-Based Approach (Preprint)

Journal of Medical Internet Research, Journal Year: 2025, Volume and Issue: 27, P. e63962 - e63962

Published: Jan. 16, 2025

Background Monitoring the emotional states of patients with psychiatric problems has always been challenging due to noncontinuous nature clinical assessments, effect health care environment, and inherent subjectivity evaluation instruments. However, mental in disorders exhibit substantial variability over time, making real-time monitoring crucial for preventing risky situations ensuring appropriate treatment. Objective This study aimed leverage new technologies deep learning techniques enable more objective, patients. was achieved by passively variables such as step count, patient location, sleep patterns using mobile devices. We predict self-reports detect sudden variations their valence, identifying that may require intervention. Methods Data this project were collected Evidence-Based Behavior (eB2) app, which records both passive self-reported daily. Passive data refer behavioral information gathered via eB2 app through sensors embedded devices wearables. These obtained from studies conducted collaboration hospitals clinics used eB2. hidden Markov models (HMMs) address missing transformer neural networks time-series forecasting. Finally, classification algorithms applied several variables, including state responses Patient Health Questionnaire-9. Results Through monitoring, we demonstrated ability accurately patients’ anticipate changes time. Specifically, our approach high accuracy (0.93) a receiver operating characteristic (ROC) area under curve (AUC) 0.98 valence classification. For predicting 1 day advance, an ROC AUC 0.87. Furthermore, feasibility forecasting Questionnaire-9, particularly strong performance certain questions. example, question 9, related suicidal ideation, model 0.9 0.77 next day’s response. Moreover, illustrated enhanced stability multivariate when HMM preprocessing combined model, opposed other methods, recurrent or long short-term memory cells. Conclusions The improved methods (eg, network memory), leveraging attention mechanisms capture longer time dependencies gain interpretability. showed potential assess scores questionnaires advance. allows hence better risk detection treatment adjustment.

Language: Английский

---

Exploring Predictors of Passive Versus Active Suicidal Ideation

Crisis, Journal Year: 2025, Volume and Issue: unknown

Published: March 18, 2025

Abstract: Background: Passive and active suicidal ideation (SI) have been shown to be co-occurring but are distinguishable constructs with presumably differential sets of predictors. Aims: The present analysis integrates nomothetic idiographic analyses unravel the relations between passive SI momentary affective states in real-time data tap several knowledge gaps. Methods: 54 psychiatric inpatients rated their current positive as well negative affect for six consecutive days (10 random prompts daily) using ecological assessments on smartphones. Data were analyzed group iterative multiple model estimation (GIMME). Results: On subgroup level, only significant contemporaneous paths emerged (with no direct from SI). In general, personalized models revealed large heterogeneity. number, direction, strengths individual differed enormously fewer than overall). interrelated majority models. Limitations: Findings limited by item wording, co-occurence SI, short observation interval. Conclusion: heterogeneous potentially reflect structural functional differences development maintenance SI.

Language: Английский

---

Integrating dynamic psychophysiological indices across time and contexts: Elucidating mechanisms, risk markers, and intervention targets

Psychophysiology, Journal Year: 2024, Volume and Issue: 61(10)

Published: July 31, 2024

Why should researchers measure psychophysiological processes repeatedly over time? The study of psychophysiology inherently involves sampling biological as they manifest time. most common approach is to use a brief sample make conclusions about how individuals or groups differ. Although these types between-subject comparisons have utility for understanding individual and group differences, many the important conceptual questions in field involve that are dynamic, varying within Using examples from literature on affect regulation, this review contrasts three designs: classic single-observation design aggregated temporally linked repeated observation designs, which great promise measuring variables fluctuate dynamically Importantly, designs can be integrated elucidate research risk (when whom will likelihood an unwanted outcome occurring increase?), mechanisms (how why does change contribute another process interest?), interventions when take place modify outcome?). Researchers encouraged implement intensive their research, conducted traditional laboratory settings (e.g., fMRI, event-related brain potentials, heart rate variability) ecologically valid contexts everyday life using ambulatory assessment.

Language: Английский

---

Mapping punishment avoidance learning deficits in non-suicidal self-injury in young adults with and without borderline personality disorder: An fMRI study

Journal of Affective Disorders, Journal Year: 2024, Volume and Issue: 370, P. 489 - 498

Published: Nov. 13, 2024

Language: Английский

---

Emotion Forecasting: A Transformer-Based Approach (Preprint)

Published: July 5, 2024

Monitoring the emotional states of psychiatric patients has always been challenging due to non-continuous nature clinical assessments, effect being in a healthcare environment, and inherent subjectivity existing evaluation instruments. However, mental disorders exhibit significant variability over time, making real-time monitoring crucial for preventing risk situations ensuring appropriate treatment. Our objective is leverage new technologies deep learning techniques enable more objective, patients. This will be achieved by passively variables like step count, patient location, sleep patterns using mobile devices. We aim predict self-reports detect sudden variations their valence, identifying that may require intervention. Data this project are registered with Evidence-Based Behavior (eB2) MindCare application, where both self-reported recorded from utilize daily summaries these variables. implement imputation methods based on hidden Markov model (HMM) address missing data transformer neural networks time-series forecasting. Finally, classification algorithms applied several variables, including state responses Patient Health Questionnaire (PHQ-9). Through monitoring, we demonstrated ability accurately state, obtaining an accuracy 0.93 0.98 receiver operating characteristic (ROC) area under curve (AUC) valence XGBoost classifier anticipate changes (ROC AUC 0.87 change detection one day advance). Additionally, showed feasibility forecasting general PHQ-9 questionnaire. Especially good results were obtained score prediction certain questions. For instance, case question 9, related suicidal ideation, 0.9 ROC 0.768 predicting following day’s response. Secondly, methodological perspective, illustrate enhanced stability multivariate when combining HMM pre-processing model, as opposed other methods, such Recurrent Neural Network or Long Short- Term Memory cells. Concretely, exploit capabilities offered attention mechanisms capture longer time dependencies. From found out improved (RNN, LSTM...), leveraging dependencies gain interpretability. show potential assess scores questionnaires passive advance. offers real hence better treatment adjustment.

Language: Английский

---